SplineResampler - possible further improvements

[MineCake147E]

Background and motivation

Currently, code size of SplineResampler is skyrocketing while it's not fast enough to handle non-integer-ratio conversion very quickly.

Possible improvements

• Reorder Catmull-Rom coefficients so that the branch will be out of dot-product loop.
  • Coefficients for CachedDirect
    • Coefficients access order will be like sawtooth wave.
  • Coefficients for CachedWrappedOdd
    • Coefficients access order will be like triangular wave.
  • Coefficients for CachedWrappedEven
    • Coefficients access order will be like triangular wave.
• Increase the maximum size of coefficient memos for wrapped strategies.
  • CachedDirect can miss the L1 cache if we increase the maximum memo size because of its sawtooth-wave-like access.
• Use delegate instead of switch statements for all combinations of channels, conversion ratio, and hardware.
  • Unify the parameter signature with newly introducing struct type.
• Optimization for the conversion ratio 44.1Hz->48kHz.
  • Conversion-ratio-specific optimizations already exist for double-rate and quadruple-rate stereo/monaural conversion.
• Utilize Expression Tree for non-trivial conversion ratio like 44.1kHz->96kHz. Currently impossible because Expression Tree supports neither ref locals nor pointers.
  • Vectorize using X86 and Arm Hardware Intrinsics if available.
  • Vectorize using System.Numerics.Vectors if no intrinsic available.
  • Unroll dot-product loops.